Presentation Information
[3GteX-05]Beyond Supervised Learning - Noise Models and Generative AI in Scientific Data Analysis
○Alexander Krull1, Benjamin Salmon2, Samuel Tonks, Thomas Muir1, Qifeng Liu1, Yehe Liu, Yuichiro Iwamoto6, Sadao Ota6, Tim-Oliver Buchholz5, Mangal Prakash4, Florian Jug3 (1. University of Birmingham (UK), 2. University of Manchester (UK), 3. Human Technopole (Italy), 4. Ellison Institute of Technology Oxford (UK), 5. Friedrich Miescher Institut (Switzerland), 6. University of Tokyo (Japan))
Keywords:
artificial intelligence,data analysis,generative AI,image restoration,denoising
Over the past decades, life sciences and biotechnology have seen a major rise in data acquisition capabilities. Modern microscopes and spectroscopic tools such as fluorescence microscopy, SMLM, transmission electron microscopy, Raman spectroscopy and flow cytometry now generate data at a scale that was unthinkable not long ago. These advances shift the central challenge from data collection to data processing. Manual approaches, for instance drawing segmentation masks by hand, were soon outpaced by the amount of data produced. Heuristic image processing, such as smoothing and thresholding, helped to some extent but was limited in accuracy and often brittle, even when designed with the characteristics of the imaging noise in mind.
Machine learning became the natural next step. Standard supervised models learn from pairs of input and target data and can perform tasks such as denoising or producing segmentation masks directly from microscopy images. These approaches mostly come from general computer vision and are adapted to biological data rather than built around domain knowledge. Their main drawback is the reliance on paired training data. Producing clean targets or segmentation masks requires considerable manual effort, and for many imaging modalities it is simply not feasible at scale.
Self supervised learning provides an attractive way around this bottleneck. Here, the idea is to draw on what we know about the physics and statistics of image formation and use this knowledge to learn from unpaired data. In the denoising setting, this means that only the noisy measurements themselves are needed. I will discuss Noise2Void as an example. It uses statistical properties of imaging noise to train models directly on raw data and still reaches results close to those of supervised models. I will also describe methods such as GAP and Bit2Bit, which rely on explicit modelling of photon statistics and again achieve self supervised denoising without requiring clean ground truth.
I will then take a Bayesian perspective on these problems. Most self supervised approaches produce a minimum mean square error estimate. This corresponds to an average over all plausible solutions and can lead to results that are overly smooth or not representative of any real underlying structure. Generative models allow us to move beyond this limitation. By modelling a conditional distribution of possible solutions, they allow us to sample individual outcomes that reflect the uncertainty in the data. These models can still be trained in a self or unsupervised manner and can incorporate knowledge about how the data are acquired.
This shift towards generative approaches raises new questions. How should we evaluate distributions instead of single reconstructions? How do we decide whether the distribution is appropriate for the task? And what does correctness mean when many solutions are possible?
Machine learning became the natural next step. Standard supervised models learn from pairs of input and target data and can perform tasks such as denoising or producing segmentation masks directly from microscopy images. These approaches mostly come from general computer vision and are adapted to biological data rather than built around domain knowledge. Their main drawback is the reliance on paired training data. Producing clean targets or segmentation masks requires considerable manual effort, and for many imaging modalities it is simply not feasible at scale.
Self supervised learning provides an attractive way around this bottleneck. Here, the idea is to draw on what we know about the physics and statistics of image formation and use this knowledge to learn from unpaired data. In the denoising setting, this means that only the noisy measurements themselves are needed. I will discuss Noise2Void as an example. It uses statistical properties of imaging noise to train models directly on raw data and still reaches results close to those of supervised models. I will also describe methods such as GAP and Bit2Bit, which rely on explicit modelling of photon statistics and again achieve self supervised denoising without requiring clean ground truth.
I will then take a Bayesian perspective on these problems. Most self supervised approaches produce a minimum mean square error estimate. This corresponds to an average over all plausible solutions and can lead to results that are overly smooth or not representative of any real underlying structure. Generative models allow us to move beyond this limitation. By modelling a conditional distribution of possible solutions, they allow us to sample individual outcomes that reflect the uncertainty in the data. These models can still be trained in a self or unsupervised manner and can incorporate knowledge about how the data are acquired.
This shift towards generative approaches raises new questions. How should we evaluate distributions instead of single reconstructions? How do we decide whether the distribution is appropriate for the task? And what does correctness mean when many solutions are possible?
Comment
To browse or post comments, you must log in.Log in